Adaptive AI-Driven Reliability Modeling: An LLM-Supported Framework for Dynamic Fitting of Failure Distributions in Safety-Critical Automotive Systems

Reliability prediction for safety-critical automotive systems traditionally relies on statistical models whose theoretical distribution functions are fixed in advance. These assumptions often fail when empirical failure data from steering electronic control units (ECUs) exhibit irregular, sparse, or non-stationary behavior, which challenges the validity of classical Software Reliability Growth Models. This paper introduces an extended reliability-modeling framework that employs Large Language Models (LLMs) to dynamically adapt and refine theoretical failure distributions. The core contribution is a reliability model in which the LLM functions as an adaptive meta-estimator: it interprets real-world failure patterns, identifies structural deviations from standard NHPP-based models, and recommends context-aware adjustments to distributional forms, parameter boundaries, and model constraints. The approach enables the automated generation of distribution-fitting strategies that incorporate both statistical evidence and domain-specific knowledge extracted from technical documentation, testing logs, and system architectures. Validation is conducted using real failure datasets from automotive steering ECUs, ensuring the method is assessed under realistic development and operational conditions. Comparative evaluations demonstrate that LLM-guided fitting improves distribution alignment, reduces systematic prediction bias, and enhances short- and mid-term forecasting accuracy, particularly for low-volume or irregular defect streams. The results indicate that integrating LLM-based meta-analysis into reliability modeling provides a scientifically grounded and practically applicable extension to traditional prediction approaches. The proposed framework supports more accurate reliability estimation and more stable parameterization.

Keywords: Software Reliability, Large Language Models, Failure Distributions, Electronic Control Units, Safety Critical Systems